Deep Learning Techniques for Power System Operation: Modeling and Implementation

The fast development of the deep learning (DL) techniques in the most recent years has drawn attention from both academia and industry. And there have been increasing applications of the DL techniques in many complex real-world situations, including computer vision, medical diagnosis, and natural language processing. The great power and flexibility of DL can be attributed to its hierarchical learning structure that automatically extract features from mass amounts of data. In addition, DL applies an end-to-end solving mechanism, and directly generates the output from the input, where the traditional machine learning methods usually break down the problem and combine the results. The end-to-end mechanism considerably improve the computational efficiency of the DL.The power system is one of the most complex artificial infrastructures, and many power system control and operation problems share the same features as the above mentioned real-world applications, such as time variability and uncertainty, partial observability, which impedes the performance of the conventional model-based methods. On the other hand, with the wide spread implementation of Advanced Metering Infrastructures (AMI), the SCADA, the Wide Area Monitoring Systems (WAMS), and many other measuring system providing massive data from the field, the data-driven deep learning technique is becoming an intriguing alternative method to enable the future development and success of the smart grid. This dissertation aims to explore the potential of utilizing the deep-learning-based approaches to solve a broad range of power system modeling and operation problems. First, a comprehensive literature review is conducted to summarize the existing applications of deep learning techniques in power system area. Second, the prospective application of deep learning techniques in several scenarios in power systems, including contingency screening, cascading outage search, multi-microgrid energy management, residential HVAC system control, and electricity market bidding are discussed in detail in the following 2-6 chapters. The problem formulation, the specific deep learning approaches in use, and the simulation results are all presented, and also compared with the currently used model-based method as a verification of the advantage of deep learning. Finally, the conclusions are provided in the last chapter, as well as the directions for future researches. It’s hoped that this dissertation can work as a single spark of fire to enlighten more innovative ideas and original studies, widening and deepening the application of deep learning technique in the field of power system, and eventually bring some positive impacts to the real-world bulk grid resilient and economic control and operation